remote sensing

Article

A Merging Framework for Rainfall Estimation at

High Spatiotemporal Resolution for Distributed
Hydrological Modeling in a Data-Scarce Area
Yinping Long 1,2 , Yaonan Zhang 1,3, * and Qimin Ma 1,2
1
2
3

Cold and Arid Regions Environmental and Engineering Research Institute of Chinese Academy of Sciences,
Lanzhou 730000, China; longyinping11@lzb.ac.cn (Y.L.), maqiming@163.com (Q.M.)
College of Resources and Environment, University of Chinese Academy of Sciences, Beijing 100049, China
Gansu Resources and Environmental Science Data Engineering Technology Research Center,
Lanzhou 730000, China
Correspondence: yaonan@lzb.ac.cn; Tel.: +86-931-4967-598

Academic Editors: Richard Gloaguen and Prasad S. Thenkabail

Received: 15 March 2016; Accepted: 5 July 2016; Published: 15 July 2016

Abstract: Merging satellite and rain gauge data by combining accurate quantitative rainfall
from stations with spatial continuous information from remote sensing observations provides a
practical method of estimating rainfall. However, generating high spatiotemporal rainfall fields
for catchment-distributed hydrological modeling is a problem when only a sparse rain gauge
network and coarse spatial resolution of satellite data are available. The objective of the study
is to present a satellite and rain gauge data-merging framework adapting for coarse resolution
and data-sparse designs. In the framework, a statistical spatial downscaling method based on the
relationships among precipitation, topographical features, and weather conditions was used to
downscale the 0.25 daily rainfall field derived from the Tropical Rainfall Measuring Mission (TRMM)
Multisatellite Precipitation Analysis (TMPA) precipitation product version 7. The nonparametric
merging technique of double kernel smoothing, adapting for data-sparse design, was combined with
the global optimization method of shuffled complex evolution, to merge the downscaled TRMM and
gauged rainfall with minimum cross-validation error. An indicator field representing the presence and
absence of rainfall was generated using the indicator kriging technique and applied to the previously
merged result to consider the spatial intermittency of daily rainfall. The framework was applied
to estimate daily precipitation at a 1 km resolution in the Qinghai Lake Basin, a data-scarce area in
the northeast of the Qinghai-Tibet Plateau. The final estimates not only captured the spatial pattern
of daily and annual precipitation with a relatively small estimation error, but also performed very
well in stream flow simulation when applied to force the geomorphology-based hydrological model
(GBHM). The proposed framework thus appears feasible for rainfall estimation at high spatiotemporal
resolution in data-scarce areas.
Keywords: downscaling; hydrological modeling; indicator kriging; merging; optimization; TRMM

1. Introduction
Accurate estimation of the amount, and the spatial and temporal distribution, of precipitation is
crucial for hydrological analysis and flood forecasting, particularly because distributed hydrological
modeling has emerged as an effective method of analyzing hydrological processes, predicting
the evolution of hydrological variables, and forecasting hydrological hazards [1,2]. A handful
of studies [36] have demonstrated that precipitation levels are the main area of uncertainty
in hydrological model predictions. Achieving reliable modeling outputs thus requires forcing
hydrological models with accurate rainfall at relative fine spatial and temporal resolutions.
Remote Sens. 2016, 8, 599; doi:10.3390/rs8070599

www.mdpi.com/journal/remotesensing

Remote Sens. 2016, 8, 599

2 of 18

Mapping rainfall is always challenging. Interpolation of point rainfall measured by rain gauges is
the conventional method (for example, with Thiessen polygons, inverse distance weighting, or kriging
techniques), but this may be subject to great uncertainty when the rain gauge network is sparse, as the
gauges can only represent rainfall information within a limited distance [7]. In contrast, remote sensing
techniques provide an evolutionary method of spatial continuous rainfall observation with a high
temporal sampling frequency. However, remote sensing products may also generate major quantitative
errors, due to cloud effects and limitations in remote sensor performance and retrieval algorithms [8].
Combining both data sources, known as data merging may, thus, be effective in maintaining both
high-quality rainfall data from stations and spatially-continuous information from remote sensing
observations [9]. Great efforts have been made to develop and evaluate algorithms for merging rain
gauge and remote sensing observations, e.g., co-kriging [10,11], linearized weighting procedures [8,12],
conditional merging [13], Barnes objective analysis [14], multi-quadric surface fitting [9], and double
kernel smoothing [15]. The outcomes of these studies are encouraging, and provide new methods of
estimating spatiotemporal precipitation, particularly in areas with limited climate data, such as the
Qinghai-Tibetan Plateau.
However, it is quite a challenging job to generate high spatial and temporal rainfall maps for
distributed hydrological modeling at a local basin scale when facing the following problems all at once:
(1) only a sparse rain gauge network is available; (2) the spatial resolution of satellite data is much
coarser than the modeled one [7]; and (3) daily rainfall is spatially intermittent [11]. This study thus
aims to present a methodology that overcomes these issues.
The kriging-based merging scheme is a common and mature spatial prediction method, but
requires the assumption of a second-order stationary and a theoretical semi-variogram model.
In poorly-gauged areas kriging-derived methods may, thus, overestimate the spatial correlation,
as distances between rain gauges are often too large and, hence, tend to deliver unsatisfactory
results [16]. Instead, nonparametric merging methods without strong spatial assumptions may be
more suitable for sparse designs. Li and Shao [15] used the nonparametric double kernel smoothing
technique to combine TRMM precipitation data with observations from the Australian rain gauge
network, focusing on discontinuity correction and spatial interpolation adapting for sparse design,
and compared this to the geostatistical methods of ordinary kriging and co-kriging. Nerini et al. [16]
compared the nonparametric rainfall methods of double kernel smoothing and mean bias correction
with two geostatistical methodskriging with external drift and Bayesian combinationfor merging
daily TRMM precipitation with rain gauge data over a mesoscale tropical Andean catchment in Peru.
Both studies concluded that the nonparametric double kernel smoothing merging method performed
better than the more complex geostatistical methods under data-scarce conditions.
Nonetheless, the precipitation with finer spatial resolution necessary for hydrological models
is still not present, as the merged results often retain the same scale as the satellite data.
Spatial downscaling of the satellite observations is thus necessary before the merging, which is a
technique for disaggregating coarse-resolution data and capturing the sub-grid heterogeneity. It is
usually based on the concept of scale invariance, or relates the properties of the physical process at
one scale to those at a finer scale [17]. A common and simple downscaling method is to develop a
statistical model at the original coarse scale, based on the relationships between rainfall and the main
factors that govern the rainfall spatial variability, and then transfer the model to the target scale, such
as the works of Jia et al. [18], Fang et al. [19], and Shi et al. [20].
Spatial intermittency means daily rainfall is delivered in discrete patches in space and time, which
causes a discontinuous surface with areas of zero rainfall between areas of non-zero rainfall [11].
However, most studies have neglected this feature, except the works of Barancourt et al. [21], Grimes
and Pardo-Igzquiza [22], and Chappell et al. [11], who tackled the discontinuity by thresholding the
rainfall distribution with the indicator kriging to map the presence and absence of rainfall.
Hence, this study presents a framework for estimating precipitation with high spatial and
temporal resolution for distributed hydrological modeling in the Qinghai Lake Basin, a data-scarce area

Remote
Remote Sens.
Sens. 2016,
2016, 8,
8, 599

3 of 18

area in the northeast of the Qinghai-Tibet Plateau. Using the 0.25 TRMM and a sparse rain gauge
in the northeast of the Qinghai-Tibet Plateau. Using the 0.25 TRMM and a sparse rain gauge network,
network, statistical spatial downscaling, double kernel smoothing merging, and indicator kriging
statistical spatial downscaling, double kernel smoothing merging, and indicator kriging techniques are
techniques are combined, following Fang et al. [19], Li and Shao [15], and Chappell et al. [11],
combined, following Fang et al. [19], Li and Shao [15], and Chappell et al. [11], respectively, to solve
respectively, to solve the issues proposed previously.
the issues proposed previously.
2. Materials and Methods
2. Materials and Methods
2.1. Study
Study Area
Area and
and Data
Data
2.1.
2.1.1. The Qinghai Lake Basin
Qinghai Lake,
Lake,the
thelargest
largest
inland
saline
in China,
in a intermountain
closed intermountain
basin
Qinghai
inland
saline
lakelake
in China,
lies inlies
a closed
basin between
2, and the

1
2
between
36323715N
and
993610047E
(Figure
1).
It
has
a
drainage
area
of
29,691
km
36 32 37 15 N and 99 36 100 47 E (Figure 1). It has a drainage area of 29,691 km , and the altitude
altitude
of the
basinfrom
ranges
from
3167
(the elevation
at the bottom
of theto
lake)
5279
m above
sea
of
the basin
ranges
3167
(the
elevation
at the bottom
of the lake)
5279tom
above
sea level
level
(asl).
Land
cover
is
dominated
by
alpine
meadow,
bare
soil,
everglade,
tundra,
and
alpine
(asl). Land cover is dominated by alpine meadow, bare soil, everglade, tundra, and alpine desert.
desert.
The primary
soilinclude
types include
felty,
thinfelty,
darkcastanozems,
felty, castanozems,
peatyand
bog,
andfrigid
dark calcic
frigid
The
primary
soil types
felty, thin
dark
peaty bog,
dark
calcic Over
soils. 40
Over
40 rivers
drain
into Qinghai
Lake,
butare
most
are intermittent.
Only
two largest
soils.
rivers
drain into
Qinghai
Lake, but
most
intermittent.
Only the
twothe
largest
rivers,
rivers,
the
Buha
and
the
Ikeulan,
have
longstanding
and
continuous
hydrological
records.
The
Buha
the Buha and the Ikeulan, have longstanding and continuous hydrological records. The Buha River
River
is the largest,
contributing
the lakes
total runoff.
The basin
is dominated
by a
is
the largest,
contributing
almostalmost
half ofhalf
theof
lakes
total runoff.
The basin
is dominated
by a cold
coldsemi-arid
and semi-arid
and influenced
three different
systems:
the East
Asian
and
climateclimate
and influenced
by threeby
different
monsoonmonsoon
systems: the
East Asian
monsoon,
monsoon,
Indian and
monsoon,
andjetwesterly
streams,
which
make
it one
the mostregions
sensitive
the
Indian the
monsoon,
westerly
streams,jetwhich
make
it one
of the
mostofsensitive
to
regionsclimate
to global
climate
[23].level
The water
leveldecreased
of the lake
decreased
overall
byfrom
about1959
3.7 to
m
global
change
[23].change
The water
of the lake
overall
by about
3.7 m
from
1959
to
2004,
but
increased
by
almost
1.5
m
from
2004
to
2013
according
to
the
observational
2004, but increased by almost 1.5 m from 2004 to 2013 according to the observational records of the
records of the
and Water
Resources
Survey
Bureau
of Qinghai
Causes
Hydrology
andHydrology
Water Resources
Survey
Bureau of
Qinghai
Province,
China.Province,
Causes ofChina.
this trend
are
of this trend
are uncertain.
Therefore,
a study of processes
the hydrological
processes
of the basin
necessary,
uncertain.
Therefore,
a study
of the hydrological
of the basin
is necessary,
with is
the
accurate
with the accurate
estimation
of precipitation
levels being one prerequisite.
estimation
of precipitation
levels
being one prerequisite.

Figure 1. Map of the Qinghai Lake Basin, showing the regional topography, main river network, rain
gauges, and discharge
discharge station
station at
at Buhahekou.
Buhahekou.

Remote Sens. 2016, 8, 599

4 of 18

2.1.2. Datasets
The rain gauge data was obtained from the China Meteorological Administration. The daily
observation is interpreted as the 24-h accumulated rainfall to 8 p.m. Beijing time (UTC + 8) for a
given day. Due to the cold alpine environment, most parts of the basin are off the beaten track and
stations are hard to maintain; as such, there are only two national weather stations within it. Thus, data
from the two stations within the basin, and eleven more stations surrounding the basin, as shown in
Figure 1, were used in the merging process. Climate factors, including maximum temperature, relative
humidity, and wind direction, which are required by the downscaling process, were also take from the
thirteen stations.
The satellite-based product was obtained from the Tropical Rainfall Measuring Mission (TRMM)
Multi-satellite Precipitation Analysis (TMPA, also TRMM 3B42) precipitation product version 7, with a
spatial resolution of 0.25 covering the latitude band 50 NS and with a temporal resolution of 3-h
covering 1998 to the present. TRMM products have been used extensively over the Qinghai-Tibetan
Plateau (e.g., [24,25]) but, to our knowledge, no merging products combine TRMM and gauge
observations from the plateau, let alone products for hydrological modeling. The TMPA products
have been scaled by monthly rain gauge analyses from the Global Precipitation Climatology Project
(GPCP) and the Climate Assessment and Monitoring System (CAMS) [26], but gauges with restricted
data accessibility cannot be used by these procedures, as is often the case in developing countries [16].
Gao and Liu [27] evaluated the performance of the TMPA version 6 products at daily scale over
the plateau, and found that light rainfall (010 mm) was overestimated and moderate and heavy
rainfall (>10 mm) were underestimated. Therefore, the TMPA products still require local adjustments.
These TMPA precipitation estimates were geographically abstracted for covering the spatial extent of
the rain gauges. To match the time-scale of the rain gauges, the satellite-based daily precipitation was
constructed by accumulating the rain falling in a 24 h period up until 8 pm Beijing time.
The Digital Elevation Model (DEM) used to derive topographical features is the Shuttle Radar
Topography Mission (SRTM) version 4, with data voids filled, available at a spatial resolution of
3 arc seconds [28]. For subsequent hydrological modeling and analysis, an equal area projection of
the datasets is required. In this study, the coordinate systems of the weather stations, satellite data,
and DEM was converted to the Albers equal area projection with GCS_Krasovsky_1940 Datum.
2.2. Merging Framework
The whole procedure can be divided into five steps (see Figure 2). First, the TRMM-derived daily
precipitation was disaggregated to a 1-by-1 km field, using a regression model developed from the
relationships among precipitation, topographical features, and weather conditions. Second, a function
optimization method known as shuffled complex evolution (SCE), developed by Duan et al. [29],
was used to automatically select two optimal bandwidths of the double kernel smoothing model.
Third, the downscaled precipitation and rain gauge observations were merged, using the double
kernel smoothing model with optimal bandwidths. Fourth, the final rainfall field was estimated by
multiplying the merged field by an indicator field, which represents the presence and absence of
rainfall, to take into account the spatial intermittency of the rainfall field. Finally, the results were
examined according to a set of performance indicators and hydrological evaluation.

Remote Sens. 2016, 8, 599

5 of 18

Remote Sens. 2016, 8, 599

5 of 18

Figure 2. Schematic overview of the rainfall estimation process.

Figure 2. Schematic overview of the rainfall estimation process.

2.2.1. Spatial Downscaling of TRMM

2.2.1. Spatial Downscaling of TRMM
Similar to the method of Fang et al. [19], the downscaling method used in this study is based on
Similar
to the that
method
Fangvariability
et al. [19], of
thedaily
downscaling
method
used
in this study
based on
the assumption
the of
spatial
precipitation
is well
captured
by theis TRMM
theproducts
assumption
that
the
spatial
variability
of
daily
precipitation
is
well
captured
by
the
TRMM
and can be explained by local topography and related meteorological conditions.products
Three
and
can be explained
by including
local topography
related
meteorologicalruggedness
conditions.index
Three
topographical
topographical
factors,
elevationand
(ELE),
the topographical
(TRI),
and the
factors,
elevation
index(ASA),
(TRI), and the
between the
angleincluding
between the
slope (ELE),
aspect the
andtopographical
the prevailingruggedness
wind direction
twoangle
meteorological
slope
aspect and
the prevailing
wind direction
(ASA), and(TEM)
two meteorological
variables,(HUM),
including
daily
variables,
including
daily maximum
air temperature
and average humidity
were
considered
the construction
theaverage
regression
model. (HUM),
These five
factors
were prepared
at 25 km
maximum
air in
temperature
(TEM)ofand
humidity
were
considered
in the construction
andregression
1 km resolutions.
of the
model. These five factors were prepared at 25 km and 1 km resolutions.
The
projectedDEM
DEMatat100
100mmresolution
resolution was
was aggregated
aggregated by
of of
The
projected
bythe
themean
meanstatistics
statisticstotoresolutions
resolutions
25
km
and
1
km.
The
resampled
DEMs
were
then
used
to
derive
the
TRI
and
slope
aspect.
The
TRI,
25 km and 1 km. The resampled DEMs were then used to derive the TRI and slope aspect. The TRI,
developed
Riley
[30]
a rapid,
objectivemeasure
measureofoftopographic
topographicheterogeneity,
heterogeneity,calculated
calculated as
developed
byby
Riley
et et
al.al.
[30]
is is
a rapid,
objective
the square
root
the averaged
squared
differences
in elevation
values
a center
theassquare
root of
theofaveraged
squared
differences
in elevation
values
fromfrom
a center
gridgrid
cellcell
andand
eight
eight
neighboring
cells.
Areas
with
high
TRI
values
indicate
mountainous
regions,
where
rainfall
is
neighboring cells. Areas with high TRI values indicate mountainous regions, where rainfall is expected
expected to be higher than elsewhere [31]. The prevailing wind direction for each month was
to be higher than elsewhere [31]. The prevailing wind direction for each month was determined as that
determined as that most frequently observed at the thirteen weather stations. We could then derive
most frequently observed at the thirteen weather stations. We could then derive the monthly aspect
the monthly aspect angle, i.e., the angle between the slope aspect and the prevailing wind direction.
angle, i.e., the angle between the slope aspect and the prevailing wind direction. The spatiotemporal
The spatiotemporal variability of temperature and humidity is not as great as that for rainfall, so the
variability of temperature and humidity is not as great as that for rainfall, so the spatial distribution of
spatial distribution of the daily maximum air temperature and average humidity can be obtained
the daily maximum air temperature and average humidity can be obtained through the interpolation
through the interpolation of ground-based observations. The inverse distance weighting (IDW)
of ground-based
observations.
The proved
inversesimple
distance
method
as thisishas
method was used,
as this has been
andweighting
suitable for(IDW)
this study
areawas
[32].used,
The method
been
proved simple
for this intensive,
study area
[32].estimates
The method
is straightforward
straightforward
andand
not suitable
computationally
which
the unknown
value at oneand
pointnot
computationally
intensive,
which
estimates
the
unknown
value
at
one
point
using
a
linearly
weighted
using a linearly weighted combination of its neighbor sample points, with the weights inversely
combination
of its
neighbor
sample
with
thethe
weights
related to the distance between
related to the
distance
between
thepoints,
estimated
and
sampleinversely
point [33].
the estimated
and the sample
point
The downscaling
process
is [33].
divided into four main steps. First, the relationships between
The downscaling
is divided
into four
main steps.
the TRI,
relationships
between
precipitation
and the process
five explanatory
variables
mentioned
above,First,
i.e., ELE,
ASA, TEM,
and
precipitation
and
the
five
explanatory
variables
mentioned
above,
i.e.,
ELE,
TRI,
ASA,
TEM,
and
HUM, were examined daily at the original scale using an univariate regression analysis, and HUM,
the
were examined daily at the original scale using an univariate regression analysis, and the final

Remote Sens. 2016, 8, 599

6 of 18

multivariate regression model was constructed at the 25 km resolution using the variables that passed
the significance test (at the p < 0.05 level). The residuals were then calculated as the differences
between the TRMM rain and the predicted values from the regression model, interpreted as the
natural random variations of precipitation not represented by the model, and interpolated to 1 km
resolution using the spline interpolator. Then, assuming that the same responses existed at the
target resolution, the regression model was applied to the estimation of 1 km precipitation using the
significant topographical and meteorological variables at the 1 km resolution. Finally, the downscaling
result was calculated by combining the interpolated residuals and the predicted precipitation.
2.2.2. The Double Kernel Smoothing Technique
The double kernel smoothing technique (DS) [15] is a nonparametric residual-based merging
method adapting to data-sparse design, which can preserve the main pattern of the satellite field if no
rain gauges are locally presented, or otherwise adjust the satellite field by using nearby observations.
The main aim is to estimate the residual field by a weighted average of point residuals, using kernel
functions, and then adjust the satellite field by the predicted residual field. It includes four steps [15]:
1.

Estimating the point residuals at each gauge location si :

Dpsi q XB psi q XO psi q

2.

where XB , XO , and D denote the background (satellite), observed, and residual fields, respectively.

Performing a first level interpolation of point residuals to generate gridded pseudo-residuals D

with a grid size of 25 km:
n

q i 1
Dps
k

K1 p||sk si ||{h1 qDpsi q

n

i 1

3.

(1)

(2)
K1 p||sk si ||{h1 q

where K is a kernel function satisfying (i) K(u) 0; (ii) K(u) = K(u); and (iii) K(u)du = 1, || || is
the Euclidean distance, and h is the bandwidth.
Applying a second level of interpolation on both the observed point residuals D and the
to generate the error field B on each grid of the 1 km downscaled
gridded pseudo-residuals D
satellite data:
n

B psq

K2 p||s si ||{h2 qDpsi q `

i 1
n

i 1

K2 p||s si ||{h2 q `

n1

k1
n1

k1

q
K2 p||s sk ||{h2 q Dps
k
(3)
K2 p||s sk ||{h2 q

As Gasser et al. [34] indicated, most kernel functions perform similarly so, in this study, the K1
and K2 kernel functions are defined as Gaussian kernels, following Nerini et al. [16]:
1
1
Kpuq ? expp u2 q
2
2
4.

(4)

Estimating the merged field XM by subtracting the error field B from the background field XB :
X M psq XB psq B psq

(5)

Remote Sens. 2016, 8, 599

7 of 18

2.2.3. Bandwidth Estimation

The bandwidth rescales the spread of the kernel function and determines the smoothness of the
estimated field [15]. An optimal bandwidth balances the bias and variance. If the bandwidth is small,
the estimate is largely influenced by observations close to the estimated point, and the estimated field
is, hence, rough with small bias, but large variance. In contrast, as the bandwidth tends to the infinite,
the estimate approaches the mean of the observations, and a large bandwidth leads to a smooth
surface with small variance, but large bias. In this study, the bandwidths h1 and h2 in steps 2 and 3
were automatically determined using the shuffled complex evolution (SCE) method developed by
Duan et al. [29], which is a global optimization strategy combining the concepts of controlled random
search, competitive evolution, and complex shuffling. The objective function of the SCE method is
based on cross-validation, given by:
CV ph1 , h2 q

D psi q B,i

(2

(6)

i 1

where B,i is the residual estimated at gauge location si by Equation (3) without using the point
residual D(si ).
2.2.4. Accounting for Spatial Intermittency
Unlike monthly or annual rainfall patterns, a daily rainfall field is usually spatially intermittent;
the rainfall surface is discontinuous, with areas of zero rainfall between areas of non-zero rainfall [11].
Most of the merged rainfall fields will be continuous, as the Gaussian kernel in the double kernel
smoothing model is a continuous function that supports infinity, which will generate a continuous
residual field. Therefore, to consider rainfall spatial intermittency, an indicator field representing the
presence and absence of rainfall was generated using the indicator kriging technique, and was applied
to the daily DS merged results. The rainfall spatial intermittency treatment is summarized below:
1.

2.
3.

4.
5.

Convert the 25 km gridded daily TRMM rainfall values into point features and append the rain
gauges with their rainfall values to the TRMM-derived point file. This ensures TRMM and gauge
rainfall values are considered in the indicator field generation.
Transform the rainfall values generated from the previous step to create a binary variable
indicating where the rainfall value is zero (0) or nonzero (1).
Under the assumption that the binary variable is stationary and autocorrelated, generate a soft
indicator field at 1 km resolution, presenting the probability of rainfall occurrence by using
ordinary kriging with a Gaussion model fitting its semi-variogram.
Produce a hard indicator field by assigning a probability threshold to the soft indicator field
(0.5 in this study).
Estimate the final rainfall field by multiplying the merged rainfall field by the hard indicator field.

2.3. Evaluation Statistics

2.3.1. Performance Indicators
To assess the performance of the merging scheme, a set of performance indicators were
examined, including the mean error (ME), percent bias (PBIAS), root mean square error (RMSE),
and Nash-Sutcliffe efficiency (NSE), respectively defined by the following equations:
n

ME

1
p p i pi q
n

(7)

i 1

PBI AS ME{p 100%

(8)

Remote Sens. 2016, 8, 599

8 of 18

g
f n
f1
p p i pi q2
RMSE e
n

(9)

i1

and

NSE 1

i1
n

p p i pi q2
(10)
ppi pq2

i 1

where n is the total number of observations, p and p are the estimated and observed values at rain
gauges, respectively, and the over bar denotes the mean value. The ME and RMSE indicate error in the
units used in the measurement. The PBIAS measures the average tendency of the estimated data to be
larger or smaller than their observed counterparts. The NSE is the relative magnitude of the residual
variance compared to the measured data variance [35]. If the overall performance is evaluated, n is the
total number of dates and stations, i.e., 366 d (for year 2008) multiplied by 13 stations in this study.
If the performance of a particular day is evaluated, n is the total number of stations, i.e., 13 stations in
this study.
2.3.2. Hydrological Evaluation
Assessment of the merging results was supplemented through distributed hydrological modeling.
The merged precipitation product was used to force the geomorphology-based hydrological model
(GBHM) [3638] for stream flow simulation, which is a physically-based distributed model that
implements very few empirical parameters and does not totally rely on the observed data for parameter
calibration and is, thus, quite useful in data-poor environments. See Yang [38] for details of the model.
The simulated daily stream flow for 2008 was compared with the observed discharge at the Buhahekou
hydrometric station (see Figure 1), which is the outlet of the largest subcatchment of the Qinghai
Lake Basin, and contributes almost half of the lakes inflow. The stream flow data was obtained from
the Hydrology and Water Resources Survey Bureau of Qinghai Province. The goodness-of-fit was
evaluated in terms of the Nash-Sutcliffe efficiency (NSE), coefficient of determination (R2 ) [39], percent
bias (PBIAS), and ratio of the root mean square error (RSR) [35] at the daily scale. Model performance
can be evaluated as very good if 0.75 < NSE < 1.0, PBIAS < 10%, and 0 RSR 0.50 [35]. The data
and parameter sets used to setup the model are listed in Table 1.
Table 1. Data and parameter sets used in the hydrological simulation.
Item

Source

DEM
Landuse
Soil parameters
LAI
NDVI

STRM V4.0 [28]

the 1:100,000 land use map of China [40]
the 1:1,000,000 soil database of China [41]
MODIS MCD15A3 [42]
MODIS MOD13Q1 [43]

3. Results
3.1. Merging Process
The rainfall estimation on 30 July 2008, a day of heavy rain before the river reached its peak flow,
is presented in this section as an example to illustrate the merging process. The derived daily TRMM
rainfall at 25 km resolution is shown in Figure 3a, and indicates that precipitation was distributed
in most of the study area, and was particularly heavy in the northeast and southeast. The rainfall
amount was significantly underestimated by TRMM when compared to the gauged rainfall. For the
downscaling process, elevation and maximum air temperature passed the initial significance test and
were used to construct the multivariate regression model (Table 2). The downscaled TRMM rainfall

Remote Sens. 2016, 8, 599

9 of 18

Remote Sens. 2016, 8, 599

9 of 18

test and were used to construct the multivariate regression model (Table 2). The downscaled TRMM
rainfall at 1 km resolution is shown in Figure 3b, and maintains the spatial pattern of the
precipitation
at theisoriginal
scale.
The3b,
optimal
bandwidths
thepattern
minimum
value
of the objective
at 1 km resolution
shown in
Figure
and maintains
the and
spatial
of the
precipitation
at the
function
achieved
the SCE
optimization
are listed
in Table
The merged
rainfall
from
original scale.
Theby
optimal
bandwidths
andmethod
the minimum
value
of the 2.
objective
function
achieved
the
downscaled
TRMM
and
the
gauged
rainfall
using
the
double
kernel
smoothing
model
with
by the SCE optimization method are listed in Table 2. The merged rainfall from the downscaled
optimal
bandwidths
shownusing
in Figure
3c. kernel
The amount
of DS
merged
rainfallbandwidths
significantly
TRMM and
the gaugedisrainfall
the double
smoothing
model
with optimal
is
increased,
the spatial
averaged
of the rainfall
Qinghaisignificantly
Lake Basinincreased,
was morethe
than
six times
the
shown in Figure
3c. The
amount rainfall
of DS merged
spatial
averaged
amount
of the
theQinghai
original Lake
TRMM,
and
themore
estimation
error
was
Table
3). However,
the
rainfall of
Basin
was
than six
times
thereduced
amount(see
of the
original
TRMM, and
spatial
intermittency
was
not preserved,
the non-rainfall
areaintermittency
disappeared,was
which
the
the estimation
error was
reduced
(see Tablei.e.,
3). However,
the spatial
not means
preserved,
rainfall
amount
of
these
areas
was
probably
overestimated.
The
DS
merged
rainfall
field
was
then
i.e., the non-rainfall area disappeared, which means the rainfall amount of these areas was probably
conditioned
by The
the DS
indicator
field
to consider
rainfall
spatial intermittency,
by setting
rainfall
overestimated.
merged
rainfall
field was
then conditioned
by the indicator
field the
to consider
amount
to zerointermittency,
in areas with by
a rainfall
of less
The
final rainfall
field
rainfall spatial
setting occurrence
the rainfall probability
amount to zero
in than
areas0.5.
with
a rainfall
occurrence
is
shown
in
Figure
3d,
and
maintains
a
rainfall
spatial
pattern
similar
to
the
original
TRMM.
The
probability of less than 0.5. The final rainfall field is shown in Figure 3d, and maintains a rainfall
evaluation
statistics
listed
in original
Table 3 TRMM.
also indicate
that the performance
of the
final rainfall
field is
spatial pattern
similar
to the
The evaluation
statistics listed
in Table
3 also indicate
slightly
than the
result.field is slightly better than the DS merged result.
that the better
performance
ofDS
themerged
final rainfall

Figure 3.
3. Rainfall
Rainfallmerging
merging
process
30 2008,
July with
2008,rainfall
with rainfall
(mm) at
gauges
process
for for
datedate
30 July
amount amount
(mm) at gauges
annotated:
annotated:
(a) The
TRMM
rainfall
at 25 km
(b) the spatial
downscaled
(a) The TRMM
derived
dailyderived
rainfalldaily
field at
25 kmfield
resolution;
(b)resolution;
the spatial downscaled
TRMM
rainfall
TRMM
field at 1(c)km
(c) the
merged
rainfall
field from
the downscaled
TRMM
field at 1rainfall
km resolution;
theresolution;
merged rainfall
field
from the
downscaled
TRMM
and gauged rainfall;
and (d)
gauged
rainfall;
and (d)
the final rainfall
indicator
conditioned
field. The
gray colorareas,
indicates
the final
indicator
conditioned
field.
The gray rainfall
color indicates
non-rainfall
and
non-rainfall
areas,
and
the
rainfall
amount
at
weather
stations
are
annotated.
the rainfall amount at weather stations are annotated.

Remote Sens. 2016, 8, 599

10 of 18

Table 2. Parameters of the main subprocesses for rainfall estimation on 30 July 2008.

Downscaling

Regression Model

r2

y = 0.001ELE + 1.042TEM 13.230

0.173

<0.05

h1 (km)

h2 (km)

CV (mm)

30.609

548.255

1596.1

Range (km)

Nugget

Sill

81.953

0.101

0.048

DS Merging

Indicator Kriging

Table 3. Evaluation statistics of the intermediate and final results for rainfall estimation on 30 July 2008.
Spatial Averaged Precipitation (mm)

ME

PBIAS

RMSE

NSE

1.90
1.89
12.54
11.40

10.99
10.22
0.46
0.44

58.05%
53.99%
2.42%
2.33%

15.32
14.52
10.34
9.90

1.11
0.89
0.04
0.12

Original TRMM
Downscaled TRMM
Merged Rainfall
Final Rainfall

3.2. Performance of Estimation

3.2.1. Overall Performance
A spatial analysis of the annual precipitation aggregated from daily 2008 estimates across the
original 25 km TRMM, spatially-downscaled TRMM, DS merged, and the final indicator conditioned
rainfall, and the PBIAS against rain gauge time series, are presented in Figure 4. The downscaled
TRMM effectively captured the spatial trends of the original TRMM, showing considerably more
rainfall distributed in the northeastern and southeastern parts of this area. The final product retained a
higher degree of information from the original TRMM than from the DS merged product, which was
reasonable given that the final merged product was conditioned by the original TRMM-derived
indicator field. Although the DS merging process significantly reduced the estimation errors,
the rainfall amount was overestimated at most weather stations. The final product shows a smaller
estimation error in the northeast where weather station density is relatively higher, and the absolute
values of PBIAS for all stations were controlled within 30%.
The daily time series of the mean error for the intermediate and final merged products are
shown in Figure 5. The original TRMM shows a poor performance between May and September,
the rainy period of this area, accounting for nearly 90% of the annual precipitation. For most days,
the amount of precipitation was underestimated by the original TRMM. The downscaled TRMM
achieved a similar performance as the original TRMM, and did not cause a higher estimation error
than the original TRMM. The mean errors was significantly reduced after the DS merging process,
but the rainfall amount was slightly overestimated for most days. The final estimation also shows
a smaller estimation error than the original TRMM. The summary of performance scores shown
in Table 4 also indicates improvements in the DS merged and final estimation product compared
with the original and downscaled TRMM. From Figure 5 and Table 4, we can also observe that the
final indicator conditioned estimation shows a slight decrease in performance compared with the
DS merged estimation, indicating that the indicator conditioning process can introduce uncertainties.
The accuracy of the final estimates is indeed directly affected by the probability threshold determining
the borders between rainfall and non-rainfall areas. A practicable way to reduce the final estimation
error is selecting a proper threshold and including more site observations.

Remote Sens. 2016, 8, 599

Remote Sens. 2016, 8, 599

11 of 18
11 of 18

Figure
4. Annual
precipitation(mm)
(mm)for
for2008
2008and
and the
the PBIAS between
gauge
Figure
4. Annual
precipitation
betweenmerging
mergingresults
resultsand
andrain
rain
gauge
time
series:
(a)
the
original
TRMM
at
25
km
resolution;
(b)
the
spatial
downscaled
TRMM
at
1 km
time series: (a) the original TRMM at 25 km resolution;
the spatial downscaled TRMM at
1 km
resolution;
Merged
result;and
and(d)
(d)the
thefinal
finalindicator
indicator conditioned
conditioned estimates.
resolution;
(c)(c)
thethe
DSDS
Merged
result;
estimates.

Remote Sens. 2016, 8, 599

12 of 18

Remote Sens. 2016, 8, 599

12 of 18

Figure 5. Comparison of daily mean errors for (a) spatial downscaled TRMM; (b) DS merged rainfall;
Figure 5. Comparison of daily mean errors for (a) spatial downscaled TRMM; (b) DS merged rainfall;
and (c) the final indicator conditioned rainfall
and (c) the final indicator conditioned rainfall

Table 4.
Evaluation statistics for rainfall estimation based on observations at the 13 stations, 2008.
Table 4. Evaluation statistics for rainfall estimation based on observations at the 13 stations, 2008.
Original TRMM
Original TRMM
Downscaled TRMM
Downscaled TRMM
Merged
Rainfall
Merged
Rainfall
Rainfall
FinalFinal
Rainfall

Spatial
Average
Precipitation
(mm)
Spatial
Average
Precipitation
(mm)
305.39
305.39
332.95
332.95
416.20
416.20
330.27
330.27

ME
ME
0.13
0.13
0.09
0.09
0.08
0.08
0.07
0.07

PBIAS
PBIAS
12.64%
12.64%
9.23%
9.23%
8.44%
8.44%
6.72%
6.72%

RMSE
RMSE
2.89
2.89
2.71
2.71
1.71
1.71
1.88
1.88

NSE
NSE
0.13
0.13
0.24
0.24
0.70
0.70
0.63
0.63

3.2.2. Performance at the Tianjun Station

3.2.2. Performance at the Tianjun Station

The evaluation of the performances in the previous section may have some limitations, as all of

The
evaluation of the performances in the previous section may have some limitations, as all of
the rain gauge observations were involved in both the merging and the validation procedure. Thus,
the rain
gauge observations
were involved
in both the
theDS
merging
and
the validation
Thus,
a cross-validation
was supplemented.
However,
merging
is time
consuming procedure.
as it combines
a cross-validation
was
supplemented.
However,
the
DS
merging
is
time
consuming
as
it
combines
the SCE method to automatically select the optimal bandwidths. The cross-validation was, thus, just
the SCE
method
selectlocated
the optimal
The (see
cross-validation
thus, just
carried
out to
at automatically
the Tianjun station,
in the bandwidths.
heart of the basin
Figure 1), by was,
performing
the Tianjun
station
without
observed
at this 1),
station.
The result (see
carriedrainfall
out at estimation
the Tianjunatstation,
located
in the
heartthe
of the
basin value
(see Figure
by performing
rainfall
Tableat5)the
alsoTianjun
shows improvements
in the
the DS
mergedvalue
and final
estimation
The
rainfall
estimation
station without
observed
at this
station. products.
The result
(see
Table 5)
estimation
error was significantly
reducedand
by the
DSestimation
merging process,
withThe
the rainfall
rainfall estimation
amount
also shows
improvements
in the DS merged
final
products.
error was significantly reduced by the DS merging process, with the rainfall amount overestimated,
and the performance decreased after the indicator conditioning process, which are consistent with that
found from the previous section.

Remote Sens. 2016, 8, 599

13 of 18

overestimated, and the performance decreased after the indicator conditioning process, which are
consistent
with8,that
Remote
Sens. 2016,
599 found from the previous section.
13 of 18
Table 5. Evaluation statistics at the Tianjun station.
Table 5. Evaluation statistics at the Tianjun station.
ME
PBIAS
RMSE
Original TRMM
0.29 ME
29.44%
PBIAS
RMSE 2.33
Downscaled TRMM
0.17
17.09%
2.13
Original TRMM
0.29
29.44%
2.33
Merged Rainfall
0.06
6.17%
1.87
Downscaled TRMM
0.17
17.09%
2.13
Final Rainfall
18.77%
Merged Rainfall 0.19 0.06
6.17%
1.87 2.04
Final Rainfall

0.19

18.77%

2.04

NSE
0.46
0.55
0.65
0.59

NSE
0.46
0.55
0.65
0.59

3.2.3. Hydrological Evaluation

3.2.3. Hydrological Evaluation
Three hydrological simulations were conducted in year 2008. Simulation I was forced by the
Three hydrological
simulations
were
conducted
year 2008.
Simulation
I was
forcedadopted
by the
interpolated
gauged rainfall
using the
IDW
method, in
a default
rainfall
estimation
method
interpolated
gauged
rainfall
using
the IDW
method,
a default
rainfall
estimation
methodresampled
adopted by
by the model.
Simulation
II was
forced
by the
TRMM
precipitation
which
was directly
to
the
model.
Simulation
II
was
forced
by
the
TRMM
precipitation
which
was
directly
resampled
to
1
km
1 km resolution using the nearest-neighbor assignment, and simulation III forced by the final
resolution
the Other
nearest-neighbor
assignment,
and simulation
III same
forcedinby
the
finalsimulations,
estimation
estimationusing
product.
model inputs
and parameters
retained the
the
three
product.
Other modelwere
inputs
parameters
the same
in the results
three simulations,
and the
and the parameters
notand
calibrated.
The retained
hydrological
modeling
are summarized
in
parameters
not calibrated.
results are
summarized
in terms
of the
terms of thewere
performance
scoresThe
of hydrological
daily stream modeling
flow simulation
at the
Buhahekou
hydrometric
performance
of daily
stream flow
thewell
Buhahekou
hydrometric
Tableby
6
station. Tablescores
6 shows
significant
NSEsimulation
increases,atas
as PBIAS
and RSR station.
reductions
shows
significant
NSE increases,
as well
as PBIAS
and RSR
reductions extremely
by simulation
simulation
III, compared
with I and
II. Thus,
simulation
III performed
well,III,
as compared
the model
with
I and II. the
Thus,
simulation
III performed
extremely
as the
model
reproduced
the product.
stream flow
reproduced
stream
flow accurately
when
forcedwell,
by the
final
estimation
rainfall
As
accurately
when forced
by the final
product.
shown2008,
in Figure
6, an after
obvious
shown in Figure
6, an obvious
peakestimation
flow can rainfall
be observed
on 1As
August
two days
the
peak
be observed
on 1inAugust
dayssimulation
after the heavy
event
in
heavyflow
raincan
event
mentioned
Section2008,
3.1. two
Except
II, therain
other
twomentioned
simulations
Section
3.1. Except
simulation
the other
two simulations
succeeded
capturing the
conditions,
succeeded
in capturing
the II,
flood
conditions,
but simulation
III inreproduced
theflood
amount
more
but
simulation
III reproduced
thealso
amount
more accurately.
results
could
observed
accurately.
Similar
results could
be observed
around 28Similar
September
2008,
thealso
lastbe
flood
of the
around
28 September
2008, thethat,
last for
flood
of the
year,
whichthe
further
indicates that,
for heavy
rain
year, which
further indicates
heavy
rain
events,
IDW-interpolated
rainfall
amounts
events,
IDW-interpolated
amounts
tend
overestimated and
TRMM rainfall
to be
tend to the
be overestimated
andrainfall
the TRMM
rainfall
to to
bebe
underestimated.
The the
overestimation
of heavy
underestimated.
overestimation
of heavy
rainfallI by
IDW also
simulation
I obtained
rainfall by IDWThe
also
explains why
simulation
obtained
a explains
high R2, why
which
indicates
a good
2
acorrelation
high R , which
indicates
a good correlation
between
the simulated
flow, but
low
between
the simulated
and observed
stream
flow, but and
low observed
NSE and stream
large PBIAS,
which
NSE
and large
PBIAS, by
which
are easily
outliers.
the daily itstream
flow
simulations,
are easily
influenced
outliers.
Frominfluenced
the daily by
stream
flowFrom
simulations,
can be
drawn
that the
itoriginal
can be drawn
the original
product
cannottobestream
directly
applied
to stream
flow study
simulation
TRMMthat
product
cannotTRMM
be directly
applied
flow
simulation
in this
area.
in
this study area.
Interpolation
basedrain
on the
sparse
rain gauge
network isifpracticable
if heavy
rainfall
Interpolation
based
on the sparse
gauge
network
is practicable
heavy rainfall
events
are
events
are especially
considered.
The proposed
framework
be used
for rainfall
estimation
especially
considered.
The proposed
mergingmerging
framework
can becan
used
for rainfall
estimation
for
for
distributed
hydrological
modeling
in this
study
area.
distributed
hydrological
modeling
in this
study
area.

Figure 6. Daily simulated and observed discharge at the Buhahekou hydrometric station.
Figure 6. Daily simulated and observed discharge at the Buhahekou hydrometric station.

Remote Sens. 2016, 8, 599

14 of 18

Table 6. Performance scores of daily stream flow, 2008.

Simulation I
Simulation II
Simulation III

NSE

R2

PBIAS

RSR

0.33
0.37
0.82

0.84
0.42
0.83

41.02%
35.87%
0.85%

0.82
0.79
0.43

4. Discussion
The statistical spatial downscaling scheme based on the relationships among precipitation,
topographical features, and weather conditions successfully represented the spatial pattern of the
precipitation fields in the original TRMM data, and did not cause higher estimation errors than the
original TRMM. The downscaling approach was initially designed for extreme convective rainfall
events, which can have different formation mechanisms. The topographical and meteorological factors
only reflect some of the environmental effects on precipitation [19]. In addition, there was very
little precipitation from October to March (about 10% of the annual precipitation), the downscaling
approach could not cause high estimation errors during that period. The estimation errors would be
further reduced as the preliminary results were then calibrated by the subsequent merging process.
Thus, although it cannot perform well every day in rainfall spatial downscaling, the statistical spatial
downscaling scheme is still applicable in this study area.
When compared with the original and downscaled TRMM, the double kernel smoothing merged
results reduced the estimation error in terms of ME, PBIAS, RMSE, and NSE, but tended to overestimate
the spatial averaged rainfall amount, particularly for heavy rains. Li and Shao [15] assigned a
specified value to bandwidth h1 and then automatically selected bandwidth h2 , which needs some
prior knowledge and does not apply to all cases. This study employed the SCE global optimization
algorithm to automatically estimate the two bandwidths for each rainy day and, thus, can achieve
optimal calibration of the TRMM rainfall. For most days, however, the amount of precipitation was
underestimated by the original TRMM, leading to negative point residuals. The searching space of
SCE was from 25 km to the length of analysis window diagonal, and the estimated bandwidths might,
therefore, be larger than the influence distance of the weather stations, which would exaggerate the
underestimation area. This was why the spatial averaged precipitations were overestimated.
The final indicator conditioned estimates captured the spatial pattern of daily and annual
precipitation with a relatively small rainfall estimation error, and also performed very well in the
stream flow simulation when forcing the GBHM model. We are, thus, able to gain insights into the
spatial distribution of precipitation at a fine resolution in the Qinghai Lake Basin for the first time.
The annual precipitation in the northwestern part of the Qinghai Lake was observed to be significantly
less than that of the central and southeastern areas (see Figure 4), which could not be identified by the
few existing weather stations. Previous studies of water balance analysis and hydrological simulation
of the lake have to either calculate precipitation from the sparse weather stations [23] or directly use
gridded precipitation products at coarse resolutions [44]. Our resulting high spatiotemporal rainfall
dataset can, therefore, be used in subsequent hydrological analysis and distributed hydrological
modeling in this area.
As there has not been any rainfall data merging research in the Qinghai Lake Basin before, we
can only just compare the rainfall product estimated by our merging framework with the estimates
obtained by other merging techniques. Co-kriging [10], combined with indicator kriging with the same
threshold, was used, by which the performance criteria obtained at the Tianjun station were 0.49 for
ME, 49.38% for PBIAS, 2.14 for RMSE, and 0.55 for NSE. For stream flow simulation, NSE, R2 , PBIAS,
and RSR were 0.83, 0.85, 25.53%, and 0.41, respectively. These figures show larger cross-validation
error and similar performance of stream flow simulation, compared with the results obtained by the
merging framework proposed in this study. Annual precipitation shows a similar spatial pattern,
but are roughly varying, with short and straight fringes of rainfall amount classes (see Figure 7b).

Remote Sens. 2016, 8, 599

Remote Sens. 2016, 8, 599

15 of 18
15 of 18

(seeour
Figure
7b). Thus,
our merging
is more
adaptive than the
kriging-based
merging
Thus,
merging
framework
is moreframework
adaptive than
the kriging-based
merging
scheme for
rainfall
scheme for
rainfall
estimationarea.
in this data-limited area.
estimation
in this
data-limited
Leaving
aside
uncertainties
about
the satellite
and weather
station
the framework
merging
Leaving
aside
uncertainties
about
the satellite
and weather
station data,
thedata,
merging
framework
contained
three
main
sources
of
uncertainties
in
the
final
results:
the
environmental
contained three main sources of uncertainties in the final results: the environmental factors used in the
factors used
in the downscaling
process,
the
search space
of the
bandwidths
particularly,
the
downscaling
process,
the search space
of the
bandwidths
and,
particularly,
the and,
threshold
determining
threshold
determining
the
borders
between
rainfall
and
non-rainfall
areas.
On
the
other
hand,
the borders between rainfall and non-rainfall areas. On the other hand, real-time merging of rain gauge
real-time
mergingdata
of rain
gauge and
remote
sensing in
data
has become
a new perspective
and
remote sensing
has become
a new
perspective
hydrological
forecasting
[2,9,45]. Byinfar,
hydrological forecasting [2,9,45]. By far, however, the merging framework proposed in this study
however, the merging framework proposed in this study can be used only to back-analyze past rainfall
can be used only to back-analyze past rainfall events, which may strongly restrict its scope of
events, which may strongly restrict its scope of application. Nonetheless, this framework has the
application. Nonetheless, this framework has the potential to be applied in real-time, as it adapts the
potential to be applied in real-time, as it adapts the key parameters of the spatial downscaling and data
key parameters of the spatial downscaling and data merging algorithms for every time-step, instead
merging algorithms for every time-step, instead of being constant for the whole period and setting at
of being constant for the whole period and setting at their optimal values. Thus, our work still has
their
optimal values. Thus, our work still has room for improvement, including (1) taking into account
room for improvement, including (1) taking into account more variables related to the geophysical
more
variables of
related
to the geophysical
mechanisms
of precipitation
in the multivariate
regression
mechanisms
precipitation
in the multivariate
regression
model; (2) assessing
the influence
ranges
model;
(2)gauges
assessing
influencedown
ranges
rain gauges
andinnarrowing
downbandwidth
the searching
ranges
of rain
andthe
narrowing
theofsearching
ranges
the automated
selection
in process,
the automated
bandwidth
selection
process,
for
more
accurate
rainfall
estimation
and
efficient
for more accurate rainfall estimation and efficient computing; (3) finding an efficient way of
computing;
(3) finding
an efficient
way
of determining
thee.g.,
threshold
applied
the indicator
field,
e.g.,
determining
the threshold
applied
to the
indicator field,
by adapting
itsto
value
for each day
and
by ensuring
adapting that
its value
for each day
and ensuring
the proportion
rainy
areas
in the
estimates
the proportion
of rainy
areas in that
the final
estimates isofthe
same
as that
infinal
the original
is the
sameand
as that
in the original
TRMM; and
(4) making
the framework
flexibleenough
and computationally
TRMM;
(4) making
the framework
flexible
and computationally
efficient
to be run in
real-time
or near
by automatically
adapting
the key parameters
thekey
algorithms,
efficient
enough
to bereal-time,
run in real-time
or near real-time,
by automatically
adaptingofthe
parameters
recoding
the algorithms
single development
that
supports high
of the
algorithms,
recoding in
theaalgorithms
in a singleenvironment
development
environment
that performance
supports high
computing,computing,
and applying
a real-time
sensing
rainfall
product,
such as
3B42-RT,
to the
performance
and
applyingremote
a real-time
remote
sensing
rainfall
product,
such
as TMPA
3B42-RT,
product
in
real-time.
to the TMPA product in real-time.

Figure
7. 7. Annual
Annualprecipitation
precipitation
(mm)
for 2008
the proposed
framework;
Figure
(mm)
for 2008
from from
(a) the(a)
proposed
merging merging
framework;
and (b)
combined
with indicator
kriging.kriging.
andco-kriging
(b) co-kriging
combined
with indicator

Conclusions
5. 5.
Conclusions
This
studyexplored
exploreda asatellite
satelliteand
and rain
rain gauge
gauge data
data merging
This
study
merging framework
frameworkfor
forrelatively
relativelyhigh
high
spatiotemporal rainfall estimation under data scarce conditions, combining the techniques of
spatiotemporal rainfall estimation under data scarce conditions, combining the techniques of statistical
statistical spatial downscaling, double kernel smoothing, shuffled complex evolution, and indicator
spatial downscaling, double kernel smoothing, shuffled complex evolution, and indicator kriging,
kriging, so as to downscale satellite rainfall products, merge satellite and rain gauge data with
so as to downscale satellite rainfall products, merge satellite and rain gauge data with minimum

Remote Sens. 2016, 8, 599

16 of 18

cross-validation error, and consider the spatial intermittency of daily rainfall. The framework was
applied to estimate daily precipitation at a 1 km resolution in the Qinghai Lake Basin. The results
of this investigation showed that the proposed merging framework was able to estimate high
spatiotemporal rainfall from the coarse Tropical Rainfall Measuring Mission (TRMM) data and
sparse rain gauge observations with a small estimation error. Stream flow simulations based on the
geomorphology-based hydrological model showed a better performance when forcing the model with
the merging results than rainfall estimated merely from the original TRMM product or interpolated
from the sparse rain gauge data. Our work sets up an example study of high spatiotemporal rainfall
estimation that takes advantage of the strengths of both remote sensing and gauged rainfall to meet
the challenges of sparse in situ data. The obtained results can be used in subsequent hydrological
analysis and distributed hydrological modeling in the Qinghai Lake Basin. Accurate estimation of daily
precipitation at fine spatial resolutions in real-time is crucial for distributed hydrological modeling and
hazards forecasting. The accuracy, generality, flexibility, and computational efficiency of the framework
are our future concerns and, thus, future studies should improve the framework for real-time running,
and evaluate the performance of the merging framework by comparing it to other estimation schemes,
and by applying it to other data-scarce areas.
Acknowledgments: This study was funded by the National Natural Science Foundation of China under Grant Nos.
91125005, J1210003/J0109 and 41301508, and the Research Cloud of Alpine Joint Observation of Chinese Academy
of Sciences under Grant No. Y329L51001. Special thanks go to data providers. Thanks also go to Supercomputing
Environment of Chinese Academy of Sciences for providing the running environment.
Author Contributions: Yinping Long and Yaonan Zhang conceived and designed the research; Yinping Long and
Qimin Ma wrote the code, processed the data and performed the experiments; Yinping Long, Yaonan Zhang and
Qimin Ma wrote the manuscript.
Conflicts of Interest: The authors declare no conflict of interest.

References
1.

2.
3.
4.

5.

6.

7.

8.
9.

Jatho, N.; Pluntke, T.; Kurbjuhn, C.; Bernhofer, C. An approach to combine radar and gauge based rainfall
data under consideration of their qualities in low mountain ranges of saxony. Nat. Hazards Earth Syst. Sci.
2010, 10, 429446. [CrossRef]
Mercogliano, P.; Segoni, S.; Rossi, G.; Sikorsky, B. Brief communication a prototype forecasting chain for
rainfall induced shallow landslides. Nat. Hazards Earth Syst. Sci. 2013, 13, 771777. [CrossRef]
Kavetski, D.; Kuczera, G.; Franks, S.W. Bayesian analysis of input uncertainty in hydrological modeling:
2. Application. Water Resour. Res. 2006, 42. [CrossRef]
Renard, B.; Kavetski, D.; Leblois, E.; Thyer, M.; Kuczera, G.; Franks, S.W. Toward a reliable decomposition of
predictive uncertainty in hydrological modeling: Characterizing rainfall errors using conditional simulation.
Water Resour. Res. 2011, 47. [CrossRef]
Salamon, P.; Feyen, L. Assessing parameter, precipitation, and predictive uncertainty in a distributed
hydrological model using sequential data assimilation with the particle filter. J. Hydrol. 2009, 376, 428442.
[CrossRef]
Syed, T.H.; Lakshmi, V.; Paleologos, E.; Lohmann, D.; Mitchell, K.; Famiglietti, J.S. Analysis of process
controls in land surface hydrological cycle over the continental united states. J. Geophys. Res. Atmos.
2004, 109. [CrossRef]
Laiolo, P.; Gabellani, S.; Campo, L.; Silvestro, F.; Delogu, F.; Rudari, R.; Pulvirenti, L.; Boni, G.; Fascetti, F.;
Pierdicca, N.; et al. Impact of different satellite soil moisture products on the predictions of a continuous
distributed hydrological model. Int. J. Appl. Earth Obs. Geoinform. 2016, 48, 131145. [CrossRef]
Woldemeskel, F.M.; Sivakumar, B.; Sharma, A. Merging gauge and satellite rainfall with specification of
associated uncertainty across australia. J. Hydrol. 2013, 499, 167176. [CrossRef]
Martens, B.; Cabus, P.; De Jongh, I.; Verhoest, N. Merging weather radar observations with ground-based
measurements of rainfall using an adaptive multiquadric surface fitting algorithm. J. Hydrol. 2013, 500,
8496. [CrossRef]

Remote Sens. 2016, 8, 599

10.
11.

12.
13.
14.
15.
16.

17.

18.
19.

20.

21.
22.
23.
24.
25.
26.

27.
28.
29.
30.
31.
32.

17 of 18

Krajewski, W.F. Cokriging radar-rainfall and rain gage data. J. Geophys. Res. Atmos. 1987, 92, 95719580.
[CrossRef]
Chappell, A.; Renzullo, L.J.; Raupach, T.H.; Haylock, M. Evaluating geostatistical methods of blending
satellite and gauge data to estimate near real-time daily rainfall for australia. J. Hydrol. 2013, 493, 105114.
[CrossRef]
Grimes, D.; Pardo-Iguzquiza, E.; Bonifacio, R. Optimal areal rainfall estimation using raingauges and satellite
data. J. Hydrol. 1999, 222, 93108. [CrossRef]
Sinclair, S.; Pegram, G. Combining radar and rain gauge rainfall estimates using conditional merging.
Atmos. Sci. Lett. 2005, 6, 1922. [CrossRef]
Rozante, J.R.; Moreira, D.S.; de Goncalves, L.G.G.; Vila, D.A. Combining TRMM and surface observations of
precipitation: Technique and validation over south america. Weather Forecast. 2010, 25, 885894. [CrossRef]
Li, M.; Shao, Q.X. An improved statistical approach to merge satellite rainfall estimates and raingauge data.
J. Hydrol. 2010, 385, 5164. [CrossRef]
Nerini, D.; Zulkafli, Z.; Wang, L.-P.; Onof, C.; Buytaert, W.; Lavado, W.; Guyot, J.-L. A comparative analysis of
TRMM-rain gauge data merging techniques at the daily time scale for distributed rainfall-runoff modelling
applications. J. Hydrometeorol. 2015. [CrossRef]
Rahman, S.; Bagtzoglou, A.C.; Hossain, F.; Tang, L.; Yarbrough, L.D.; Easson, G. Investigating spatial
downscaling of satellite rainfall data for streamflow simulation in a medium-sized basin. J. Hydrometeorol.
2009, 10, 10631079. [CrossRef]
Jia, S.; Zhu, W.; Lu,
A.; Yan, T. A statistical spatial downscaling algorithm of TRMM precipitation based on
NDVI and dem in the qaidam basin of China. Remote Sens. Environ. 2011, 115, 30693079. [CrossRef]
Fang, J.; Du, J.; Xu, W.; Shi, P.J.; Li, M.; Ming, X.D. Spatial downscaling of TRMM precipitation data based on
the orographical effect and meteorological conditions in a mountainous area. Adv. Water Resour. 2013, 61,
4250. [CrossRef]
Shi, Y.L.; Song, L.; Xia, Z.; Lin, Y.R.; Myneni, R.B.; Choi, S.H.; Wang, L.; Ni, X.L.; Lao, C.L.; Yang, F.K.
Mapping annual precipitation across mainland China in the period 20012010 from TRMM3B43 product
using spatial downscaling approach. Remote Sens. 2015, 7, 58495878. [CrossRef]
Barancourt, C.; Creutin, J.D.; Rivoirard, J. A method for delineating and estimating rainfall fields.
Water Resour. Res. 1992, 28, 11331144. [CrossRef]
Grimes, D.I.F.; Pardo-Iguzquiza, E. Geostatistical analysis of rainfall. Geogr. Anal. 2010, 42, 136160.
[CrossRef]
Li, X.-Y.; Xu, H.-Y.; Sun, Y.-L.; Zhang, D.-S.; Yang, Z.-P. Lake-level change and water balance analysis at lake
Qinghai, west China during recent decades. Water Resour. Manag. 2007, 21, 15051516. [CrossRef]
Tong, K.; Su, F.; Yang, D.; Zhang, L.; Hao, Z. Tibetan plateau precipitation as depicted by gauge observations,
reanalyses and satellite retrievals. Int. J. Climatol. 2014, 34, 265285. [CrossRef]
Yang, K.; Guo, X.; He, J.; Qin, J.; Koike, T. On the climatology and trend of the atmospheric heat source over
the Tibetan plateau: An experiments-supported revisit. J. Clim. 2010, 24, 15251541. [CrossRef]
Huffman, G.J.; Bolvin, D.T.; Nelkin, E.J.; Wolff, D.B.; Adler, R.F.; Gu, G.; Hong, Y.; Bowman, K.P.; Stocker, E.F.
The TRMM multisatellite precipitation analysis (TMPA): Quasi-global, multiyear, combined-sensor
precipitation estimates at fine scales. J. Hydrometeorol. 2007, 8, 3855. [CrossRef]
Gao, Y.C.; Liu, M.F. Evaluation of high-resolution satellite precipitation products using rain gauge
observations over the Tibetan Plateau. Hydrol. Earth Syst. Sci. 2013, 17, 837849. [CrossRef]
Reuter, H.I.; Nelson, A.; Jarvis, A. An evaluation of void-filling interpolation methods for SRTM data. Int. J.
Geogr. Inf. Sci. 2007, 21, 9831008. [CrossRef]
Duan, Q.Y.; Gupta, V.K.; Sorooshian, S. Shuffled complex evolution approach for effective and efficient global
minimization. J. Optim. Theory Appl. 1993, 76, 501521. [CrossRef]
Riley, S.J.; DeGloria, S.D.; Elliot, R. A terrain ruggedness index that quantifies topographic heterogeneity.
Intermt. J. Sci. 1999, 5, 14.
Prudhomme, C.; Reed, D.W. Relationships between extreme daily precipitation and topography in a
mountainous region: A case study in Scotland. Int. J. Climatol. 1998, 18, 14391453. [CrossRef]
Long, Y.; Zhang, Y.; Zhao, G.; Yan, P.; Li, Q.; LI, R. The uncertainty in meteorological and hydrological
processes modeled by using swat modelA case study in the buhachu river basin. J. Glaciol. Geocryol. 2012,
34, 660667.

Remote Sens. 2016, 8, 599

33.
34.
35.
36.
37.
38.
39.
40.
41.

42.

43.

44.
45.

18 of 18

Lu, G.Y.; Wong, D.W. An adaptive inverse-distance weighting spatial interpolation technique. Comput. Geosci.
2008, 34, 10441055. [CrossRef]
Gasser, T.; Muller, H.G.; Mammitzsch, V. Kernels for nonparametric curve estimation. J. R. Stat. Soc.
B Methodol. 1985, 47, 238252.
Moriasi, D.; Arnold, J.; Van Liew, M.; Bingner, R.; Harmel, R.; Veith, T. Model evaluation guidelines for
systematic quantification of accuracy in watershed simulations. Trans. ASABE 2007, 50, 885900. [CrossRef]
Yang, D.; Oki, T.; Herath, S.; Musiake, K.; Singh, V.; Frevert, D. A geomorphology-based hydrological model
and its applications. Math. Models Small Watershed Hydrol. Appl. 2002, 259300.
Yang, D.; Herath, S.; Musiake, K. A hillslope-based hydrological model using catchment area and width
functions. Hydrol. Sci. J. 2002, 47, 4965. [CrossRef]
Yang, D. Distributed hydrologic Model Using Hillslope Discretization Based on Catchment Area Function:
Development and Applications; University of Tokyo: Tokyo, Japan, 1998.
Legates, D.R.; McCabe, G.J. Evaluating the use of goodness-of-fit measures in hydrologic and hydroclimatic
model validation. Water Resour. Res. 1999, 35, 233241. [CrossRef]
Liu, J.; Liu, M.; Zhuang, D.; Zhang, Z.; Deng, X. Study on spatial pattern of land-use change in china during
19952000. Sci. China 2003, 46, 373384.
Shi, X.; Yu, D.; Pan, X.; Sun, W.; Gong, Z.; Warner, E.; Petersen, G. A Framework for the 1:1,000,000
Soil Database of China. In Proceedings of the 17th World Congress of Soil Science, Bangkok, Thailand,
1421 August 2002.
NASA LP DAAC. Nasa Eosdis Land Processes Daac, Usgs Earth Resources Observation and Science (EROS)
Center, Sioux Falls, South Dakota. Available online: https://lpdaac.Usgs.Gov/dataset_discovery/modis/
modis_products_table/mcd15a3 (accessed on 4 November 2014).
NASA LP DAAC. Nasa Eosdis Land Processes Daac, Usgs Earth Resources Observation and Science (EROS)
Center, Sioux Falls, South Dakota. Available online: https://lpdaac.Usgs.Gov/dataset_discovery/modis/
modis_products_table/mcd13q1 (accessed on 9 January 2015).
Zhang, G.Q.; Xie, H.J.; Yao, T.D.; Li, H.Y.; Duan, S.Q. Quantitative water resources assessment of Qinghai
Lake basin using snowmelt runoff model (SRM). J. Hydrol. 2014, 519, 976987. [CrossRef]
Sideris, I.V.; Gabella, M.; Erdin, R.; Germann, U. Real-time radar-rain-gauge merging using spatio-temporal
co-kriging with external drift in the alpine terrain of Switzerland. Q. J. R. Meteorol. Soc. 2013, 140, 10971111.
[CrossRef]
2016 by the authors; licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC-BY) license (http://creativecommons.org/licenses/by/4.0/).